Dynamic consolidation of virtual machines (VMs) is a promising technology for reducing energy consumption of data centers. Existing studies on VM consolidation, however, are based on precopy live migration; it is difficult to optimize VM locations aggressively due to its long and undeterminable migration process. It is diffi.cult to optimize VM locations quickly on sudden load changes, resulting in serious violations of VM performance criteria and inefficient energy use of datacenters.
In this project, we propose an energy-efficient VM consolidation system exploiting postcopy live migration, which always allows quick live migration for any VMs. The consolidation system can optimize VM locations and server power states more frequently than those of using precopy live migration.
We developed the prototype of a consolidation system using our postcopy live migration mechanism (Yabusame), where excessive hardware nodes were suspended by means of ACPI S3 and all power usages were monitored with watt meters. VM locations are reactively optimized in response to ever-changing resource usage. Sudden overloading of server nodes are promptly resolved by quickly switching the execution hosts of VMs.
We confirmed that our consolidation system achieved a higher degree of performance assurance than using precopy migration. Our micro benchmark program, designed for the metric of performance assurance, showed that performance degradation was only 12% or less, even for memory-intensive workloads, which was less than half the level of using precopy live migration. The SPECweb benchmark showed that performance degrada-tion was approximately 10%, which was greatly alleviated from the case of using precopy live migration (21%).
Our experiments also showed that our consolidation system with postcopy live migration eliminated more excessive power consumption than that of using precopy live migration. Postcopy live migration allowed the prototype system to eliminate 11.8% energy overheads of actively-running VMs, which was improved by approximately 50% from precopy live migration.
Our consolidation system aggressively optimizes VM locations for better energy saving and performance assurance. It repacks VM locations every 10 seconds.
The below graphs show CPU usage of physical hosts and VMs as well as VM locations. The system automatically relocated VMs in response to CPU loads of VMs. When VMs are consolidated, excess energy use is eliminated by suspending unused physical hosts. When VMs are distributed, VM performance is assured.
We developed a power measuring system of our server cluster, which periodically collects power consumption of host nodes and network switches individually. The current, voltage, and active power of a target component are measured by a customized watt meter. The accuracy of active power is .2%. The measurement interval of the watt meter is one second. All watt meters are connected to a monitoring server (i.e., Power Monitor Node) via USB interfaces. It is possible to measure power consumption of 120 target components. The below is a photo of a part of our power measuring system; a 2U rackmount measuring board for 8 target components is installed into a 19-inch rack.
If you prefer publications in Japanese, you can find technical reports and talk slides here.